RSNA 2022 Cervical Spine Fracture Detection

Perform Exploratory Data Analysis

(As part of continuous learning - Akshu)

@akshuvenkataraman
We have included all the references and acknowledgements¶

Install and import required libraries¶

In [9]:
import os
for dirname, _, filenames in os.walk('/kaggle/input/wheelfiles'):
    for filename in filenames:
        print(os.path.join(dirname, filename))
        
!pip install -q /kaggle/input/wheelfiles/pylibjpeg-1.4.0-py3-none-any.whl
!pip install -q /kaggle/input/wheelfiles/python_gdcm-3.0.14-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
!pip install -q /kaggle/input/wheelfiles/pylibjpeg_libjpeg-1.3.1-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

/kaggle/input/wheelfiles/pylibjpeg-1.4.0-py3-none-any.whl
/kaggle/input/wheelfiles/efficientnet-1.1.1-py3-none-any.whl
/kaggle/input/wheelfiles/pydicom-2.3.0-py3-none-any.whl
/kaggle/input/wheelfiles/python_gdcm-3.0.16-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
/kaggle/input/wheelfiles/pylibjpeg_libjpeg-1.3.1-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
/kaggle/input/wheelfiles/efficientnetb5_notop.h5
/kaggle/input/wheelfiles/gdcm-1.1-py3-none-manylinux1_x86_64.whl
/kaggle/input/wheelfiles/Keras_Applications-1.0.8-py3-none-any.whl
/kaggle/input/wheelfiles/python_gdcm-3.0.14-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: There was an error checking the latest version of pip.
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: There was an error checking the latest version of pip.
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: There was an error checking the latest version of pip.
In [10]:
import os
import random
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

import plotly.express as px
from plotly.offline import init_notebook_mode, iplot, plot
from tqdm.notebook import tqdm
from pathlib import Path
from collections import Counter

tqdm.pandas()
In [11]:
# for dcm and nii
import pydicom
import nibabel as nib
from pydicom.pixel_data_handlers.util import apply_voi_lut

The Vertebral Column¶

Source: https://www.cancer.gov¶

The spine is made up of bones, muscles, tendons, nerves, and other tissues that reach from the base of the skull near the spinal cord (clivus) to the coccyx (tailbone). The vertebrae (back bones) of the spine include the cervical spine (C1-C7), thoracic spine (T1-T12), lumbar spine (L1-L5), sacral spine (S1-S5), and the tailbone. Each vertebra is separated by a disc. The vertebrae surround and protect the spinal cord.

Image Source

Learning More About the Cervical Spine¶

Source: https://my.clevelandclinic.org/¶

What is the cervical spine?¶

Your cervical spine — the neck area of your spine — consists of seven stacked bones called vertebrae.

Marked C1 to C7 in the above image

What does the cervical spine do?¶

  • Protecting your spinal cord.
  • Supporting your head and allowing movement.
  • Providing a safe passageway for vertebral arteries.

Cervical Spine Fracture¶

A fracture to the bones of your spine can result from compression (often from minor trauma in a person with osteoporosis) or be a burst fracture (vertebra that’s crushed in all directions) or a fracture-dislocation (mostly from vehicle accidents or falls from heights).

Tests and Imaging¶

Computed tomography (CT) scan: This scan uses X-rays and computers to produce images that are very thin “slices” of the area under examination. A CT scan can show the shape and size of your spinal canal, its contents and the bone around it. It helps diagnose bone spurs, osteophytes, bone fusion and bone destruction from infection or tumor.

Competition Goal¶

The goal of this competition is to identify fractures in CT scans of the cervical spine (neck) at both the level of a single vertebrae and the entire patient.

Quickly detecting and determining the location of any vertebral fractures is essential to prevent neurologic deterioration and paralysis after trauma.

In [12]:
paths = {
    'train_df': Path('../input/rsna-2022-cervical-spine-fracture-detection/train.csv'),
    'train_bbox': Path('../input/rsna-2022-cervical-spine-fracture-detection/train_bounding_boxes.csv'),
    'train_images': Path('../input/rsna-2022-cervical-spine-fracture-detection/train_images'),
    'train_nifti_segments': Path('../input/rsna-2022-cervical-spine-fracture-detection/segmentations'),
    'test_df': Path('../input/rsna-2022-cervical-spine-fracture-detection/test.csv'),
    'test_images': Path('../input/rsna-2022-cervical-spine-fracture-detection/test_images')
}

Preparing Main DataFrame¶

In [13]:
train_df = pd.read_csv(paths['train_df'])
test_df = pd.read_csv(paths['test_df'])

Column Information¶

  • StudyInstanceUID - The study ID. There is one unique study ID for each patient scan.
  • patient_overall - One of the target columns. The patient level outcome, i.e. if any of the vertebrae are fractured.
  • C[1-7] - The other target columns. Whether the given vertebrae is fractured. See this diagram for the real location of each vertbrae in the spine.
In [14]:
train_df.head()
Out[14]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7
0 1.2.826.0.1.3680043.6200 1 1 1 0 0 0 0 0
1 1.2.826.0.1.3680043.27262 1 0 1 0 0 0 0 0
2 1.2.826.0.1.3680043.21561 1 0 1 0 0 0 0 0
3 1.2.826.0.1.3680043.12351 0 0 0 0 0 0 0 0
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0

Ignoring a specific StudyInstanceUID¶

Discussion Link: The scan 1.2.826.0.1.3680043.20574 does not include a full cervical spine and should be ignored. Thanks to @kretes for catching this.

In [15]:
train_df[train_df['StudyInstanceUID'] == '1.2.826.0.1.3680043.20574']
Out[15]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7
1135 1.2.826.0.1.3680043.20574 1 0 1 0 0 0 0 0
In [16]:
train_df.drop(index=1135, inplace=True)
train_df.reset_index(drop=True, inplace=True)
train_df.head()
Out[16]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7
0 1.2.826.0.1.3680043.6200 1 1 1 0 0 0 0 0
1 1.2.826.0.1.3680043.27262 1 0 1 0 0 0 0 0
2 1.2.826.0.1.3680043.21561 1 0 1 0 0 0 0 0
3 1.2.826.0.1.3680043.12351 0 0 0 0 0 0 0 0
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0

adding total_fractures column to indicate how many cervical vertebrae are fractured¶

In [17]:
train_df['total_fractures'] = train_df.loc[:,[f"C{i}" for i in range(1,8)]].sum(axis=1)
In [18]:
train_df.head()
Out[18]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7 total_fractures
0 1.2.826.0.1.3680043.6200 1 1 1 0 0 0 0 0 2
1 1.2.826.0.1.3680043.27262 1 0 1 0 0 0 0 0 1
2 1.2.826.0.1.3680043.21561 1 0 1 0 0 0 0 0 1
3 1.2.826.0.1.3680043.12351 0 0 0 0 0 0 0 0 0
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0 1
In [19]:
len(train_df)
Out[19]:
2018

Now in total, We've 2018 study IDs¶

Understanding CT Scans¶

...A CT scan generates images that can be reformatted in multiple planes. It can even generate three-dimensional images...read more

This allows multiple planes of observation, the most common one being axial.

CT Scan Planes

Image Source

The data in train_images and test_images is segmented as per the axial plane.¶

Adding segmentations path of each patient to the dataframe¶

In [21]:
pd.set_option('display.max_colwidth', None)
train_df['segment_path'] = train_df['StudyInstanceUID'].map(lambda x: paths['train_images']/x)
train_df.head()
Out[21]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7 total_fractures segment_path
0 1.2.826.0.1.3680043.6200 1 1 1 0 0 0 0 0 2 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.6200
1 1.2.826.0.1.3680043.27262 1 0 1 0 0 0 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.27262
2 1.2.826.0.1.3680043.21561 1 0 1 0 0 0 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.21561
3 1.2.826.0.1.3680043.12351 0 0 0 0 0 0 0 0 0 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.12351
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.1363

Each patient has different number of slices, adding num_slices to each patient¶

In [22]:
def num_slices(path):
    slices = list(path.glob('*'))
    return len(slices)

train_df['num_slices'] = train_df['segment_path'].progress_map(num_slices)
train_df['num_slices'] = train_df['num_slices'].astype('int')

  0%|          | 0/2018 [00:00<?, ?it/s]
In [23]:
train_df.head()
Out[23]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7 total_fractures segment_path num_slices
0 1.2.826.0.1.3680043.6200 1 1 1 0 0 0 0 0 2 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.6200 243
1 1.2.826.0.1.3680043.27262 1 0 1 0 0 0 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.27262 406
2 1.2.826.0.1.3680043.21561 1 0 1 0 0 0 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.21561 385
3 1.2.826.0.1.3680043.12351 0 0 0 0 0 0 0 0 0 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.12351 501
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.1363 199

Let us review Patient fracture vs non-fracture, number of fractures distribution from training dataset¶

In [24]:
fig = px.pie(train_df['patient_overall'].value_counts().reset_index(), 
       names=['not_fractured','fractured'], # [0,1]
       values='patient_overall',
       color_discrete_sequence = px.colors.qualitative.Pastel,
       title='Patient Fracture Overall Distribution'
      ).update_traces(textinfo='label+percent')
iplot(fig)
In [25]:
fractures = train_df.loc[:,[f"C{i}" for i in range(1,8)]].sum().reset_index()
fractures['not_fractured'] = fractures[0].map(lambda x:len(train_df)-x)
fractures.rename(columns={'index':'vertebra',0:'fractured'}, inplace=True)
fig = px.bar(fractures, 
             x='vertebra', 
             y=['fractured','not_fractured'],
             title='Fractures as per Vertebra',
             color_discrete_sequence = ['#e0465f','#afde2c'],
            )
iplot(fig)
In [26]:
fig = px.bar(train_df['total_fractures'].value_counts().reset_index(), 
             x='index', 
             y='total_fractures',
             title='No. of fractures amongst patients',
             text='total_fractures',
             color_discrete_sequence = ['#d82cde'],
            )
iplot(fig)
In [27]:
fig = px.histogram(train_df, x="num_slices")
iplot(fig)

Understanding the different medical-imaging file formats¶

  • .dcm - DICOM file: A DCM file is an image file saved in the Digital Imaging and Communications in Medicine (DICOM) image format. It stores a medical image, such as a CT scan or ultrasound, and may also include patient information to pair the image with the patient.

  • .nii - NIFTI file format - A nibabel image object is the association of three things:

    • an N-D array containing the image data;
    • a (4, 4) affine matrix mapping array coordinates to coordinates in some RAS+ world coordinate space (Coordinate systems and affines);
    • image metadata in the form of a header.

Loading DICOM and NIFTI files¶

from @andradaolteanu and @harshitsheoran¶

Let's try opening a random .dcm and a .nii file¶

.dcm file¶

In [28]:
random_dcm_file = list(train_df['segment_path'][123].glob('*'))[0]
print(random_dcm_file)
random_dcm_file = pydicom.dcmread(random_dcm_file)
print(random_dcm_file)

../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.21724/257.dcm
Dataset.file_meta -------------------------------
(0002, 0001) File Meta Information Version       OB: b'\x00\x01'
(0002, 0002) Media Storage SOP Class UID         UI: CT Image Storage
(0002, 0003) Media Storage SOP Instance UID      UI: 1.2.826.0.1.3680043.21724.1.257
(0002, 0010) Transfer Syntax UID                 UI: Implicit VR Little Endian
(0002, 0012) Implementation Class UID            UI: 1.2.40.0.13.1.1.1
(0002, 0013) Implementation Version Name         SH: 'PYDICOM 2.3.0'
-------------------------------------------------
(0008, 0018) SOP Instance UID                    UI: 1.2.826.0.1.3680043.21724.1.257
(0008, 0023) Content Date                        DA: '20220727'
(0008, 0033) Content Time                        TM: '181031.112496'
(0010, 0010) Patient's Name                      PN: '21724'
(0010, 0020) Patient ID                          LO: '21724'
(0018, 0050) Slice Thickness                     DS: '0.625'
(0020, 000d) Study Instance UID                  UI: 1.2.826.0.1.3680043.21724
(0020, 000e) Series Instance UID                 UI: 1.2.826.0.1.3680043.21724.1
(0020, 0013) Instance Number                     IS: '257'
(0020, 0032) Image Position (Patient)            DS: [-87.257, -14.763, -149.388]
(0020, 0037) Image Orientation (Patient)         DS: [1.000000, 0.000000, 0.000000, 0.000000, 1.000000, 0.000000]
(0028, 0002) Samples per Pixel                   US: 1
(0028, 0004) Photometric Interpretation          CS: 'MONOCHROME2'
(0028, 0010) Rows                                US: 512
(0028, 0011) Columns                             US: 512
(0028, 0030) Pixel Spacing                       DS: [0.296875, 0.296875]
(0028, 0100) Bits Allocated                      US: 16
(0028, 0101) Bits Stored                         US: 16
(0028, 0102) High Bit                            US: 15
(0028, 0103) Pixel Representation                US: 1
(0028, 1050) Window Center                       DS: '500.0'
(0028, 1051) Window Width                        DS: '2000.0'
(0028, 1052) Rescale Intercept                   DS: '-1024.0'
(0028, 1053) Rescale Slope                       DS: '1.0'
(7fe0, 0010) Pixel Data                          OW: Array of 524288 elements
In [29]:
# getting metadata with .get(key)
print("Instance Number:",random_dcm_file.get('InstanceNumber'))
print("Rows x Columns:", random_dcm_file.get("Rows"), random_dcm_file.get("Columns"))
print("Image Position (Patient):", random_dcm_file.get("ImagePositionPatient"))

Instance Number: 257
Rows x Columns: 512 512
Image Position (Patient): [-87.257, -14.763, -149.388]

from the above metadata:¶

  • patient ID = patient Name
  • Slice Thickness = 1.0 mm [might not be same for all patients]
  • Instance Number = slice number
  • Image Position (Patient): gives an array of values [x-axis, y-axis, z-axis], here z-axis denotes the position in the sagittal plane

apply_voi_lut() applies a windowing function to the image - read more here

In [30]:
dcm_image = apply_voi_lut(random_dcm_file.pixel_array, random_dcm_file)
In [31]:
plt.imshow(dcm_image, cmap='bone')
Out[31]:
<matplotlib.image.AxesImage at 0x7f3d27ee8410>

.nii file¶

In [32]:
random_nii_file = list(paths['train_nifti_segments'].glob('*'))[10]
print(random_nii_file)

def open_nii_file(path):
    f = nib.load(path)
    segmentations = f.get_fdata()[:, ::-1, ::-1].transpose(2, 1, 0)
    return segmentations

nii_segments = open_nii_file(random_nii_file)
print(nii_segments.shape, "=> (num_slices, height, width)")
plt.imshow(nii_segments[123,:,:],cmap='bone')

../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.32436.nii
(271, 512, 512) => (num_slices, height, width)
Out[32]:
<matplotlib.image.AxesImage at 0x7f3d24624dd0>

A caveat of axial plane is that we are not aware of which cervical vertebra we are looking at, the above nii_segments array can be reused to get the vertebra index using np.unique()¶

In [33]:
# np.unique(nii_segments[slice_number])
print("[background, *, vertebra]")
print("here 4. denotes that the slice is a part of C4 vertebra")
np.unique(nii_segments[100])

[background, *, vertebra]
here 4. denotes that the slice is a part of C4 vertebra
Out[33]:
array([0., 3., 4.])

not all study UIDs have nii segments, for the ones we do have, adding their path to df¶

In [34]:
f"We only have {len(list(paths['train_nifti_segments'].glob('*')))} nii files / {len(train_df)} studies"
Out[34]:
'We only have 87 nii files / 2018 studies'
In [35]:
def add_nii_segment_path(uid):
    base_path = paths['train_nifti_segments']
    # path if exists else None
    path = base_path/(uid+'.nii')
    if path.exists():
        return path
    return None

train_df['nii_segments_path'] = train_df['StudyInstanceUID'].map(add_nii_segment_path)
train_df.head()
Out[35]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7 total_fractures segment_path num_slices nii_segments_path
0 1.2.826.0.1.3680043.6200 1 1 1 0 0 0 0 0 2 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.6200 243 None
1 1.2.826.0.1.3680043.27262 1 0 1 0 0 0 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.27262 406 None
2 1.2.826.0.1.3680043.21561 1 0 1 0 0 0 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.21561 385 None
3 1.2.826.0.1.3680043.12351 0 0 0 0 0 0 0 0 0 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.12351 501 None
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.1363 199 ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.1363.nii
In [36]:
def get_dcm_images(path):
    paths = list(path.glob('*'))
    paths.sort(key=lambda x:int(x.stem)) # sort based on slice index which is the filename: index.dcm
    data = [pydicom.dcmread(f) for f in paths]
    images = [apply_voi_lut(dcm.pixel_array, dcm) for dcm in data]
    return images

def get_nii_segments(path):
    f = nib.load(path)
    segmentations = f.get_fdata()[:, ::-1, ::-1].transpose(2, 1, 0)
    return segmentations
In [37]:
sample_idx = 99
dcm_images = get_dcm_images(train_df['segment_path'][sample_idx])
nii_segments = get_nii_segments(train_df['nii_segments_path'][sample_idx])
print((len(dcm_images), *dcm_images[0].shape), nii_segments.shape)

(271, 512, 512) (271, 512, 512)
In [38]:
def plot_slice(idx, dcm_images=dcm_images, nii_segments=nii_segments):
    fig, (ax1, ax2) = plt.subplots(1, 2)
    ax1.axis('off'); ax2.axis('off')
    fig.suptitle(f'Slice {idx}')
    ax1.imshow(dcm_images[idx], cmap='bone')
    ax2.imshow(nii_segments[idx,:,:], cmap='bone')

for i in range(123,128):
    plot_slice(i)

all UIDs which have segmentations:¶

In [39]:
uids_with_segments = train_df[train_df['nii_segments_path'].notnull()]
uids_with_segments.head()
Out[39]:
StudyInstanceUID patient_overall C1 C2 C3 C4 C5 C6 C7 total_fractures segment_path num_slices nii_segments_path
4 1.2.826.0.1.3680043.1363 1 0 0 0 0 1 0 0 1 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.1363 199 ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.1363.nii
41 1.2.826.0.1.3680043.25704 0 0 0 0 0 0 0 0 0 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.25704 257 ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.25704.nii
53 1.2.826.0.1.3680043.20647 0 0 0 0 0 0 0 0 0 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.20647 263 ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.20647.nii
60 1.2.826.0.1.3680043.31077 1 0 0 1 1 1 1 0 4 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.31077 672 ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.31077.nii
99 1.2.826.0.1.3680043.17960 0 0 0 0 0 0 0 0 0 ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.17960 271 ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.17960.nii

some scans maybe in reverse order¶

A Segmentation is in reverse order by @itsuki9180 and based on this comment by @abebe9849

based on this discussion, for the z_pos value in ImagePositionPatient. If z_pos of last slice is greater than z_pos of first slice, then we've to reverse the order since positive is towards the head of the patient¶

In [40]:
def check_reverse_required(path):
    paths = list(path.glob('*'))
    paths.sort(key=lambda x:int(x.stem))
    z_first = pydicom.dcmread(paths[0]).get("ImagePositionPatient")[-1]
    z_last = pydicom.dcmread(paths[-1]).get("ImagePositionPatient")[-1]
    if z_last < z_first:
        return False
    return True
In [41]:
checks = train_df['segment_path'].progress_map(check_reverse_required)

  0%|          | 0/2018 [00:00<?, ?it/s]
In [42]:
train_df['reverse_required'] = checks
In [43]:
Counter(checks)
Out[43]:
Counter({False: 1822, True: 196})
In [44]:
indices_where_reverse_required = [i for i,req in (checks.reset_index()).values if req is True]
print(indices_where_reverse_required)

[1, 4, 27, 30, 39, 47, 48, 66, 72, 80, 136, 151, 169, 176, 194, 195, 201, 209, 214, 232, 235, 243, 258, 262, 267, 281, 297, 332, 334, 336, 338, 339, 344, 372, 386, 391, 402, 407, 419, 421, 428, 435, 455, 467, 476, 488, 490, 502, 507, 510, 517, 534, 543, 558, 561, 567, 569, 572, 574, 588, 600, 606, 617, 633, 635, 640, 671, 680, 685, 702, 716, 721, 756, 777, 779, 787, 802, 803, 831, 832, 837, 840, 846, 870, 886, 888, 897, 906, 912, 913, 928, 937, 941, 943, 958, 978, 995, 999, 1011, 1017, 1022, 1024, 1041, 1055, 1056, 1074, 1075, 1086, 1094, 1102, 1104, 1108, 1120, 1138, 1142, 1157, 1168, 1196, 1222, 1226, 1237, 1255, 1274, 1283, 1310, 1332, 1339, 1346, 1382, 1383, 1393, 1395, 1407, 1408, 1412, 1431, 1437, 1444, 1467, 1481, 1484, 1486, 1487, 1494, 1496, 1497, 1504, 1509, 1513, 1524, 1529, 1530, 1531, 1538, 1554, 1566, 1583, 1587, 1594, 1607, 1614, 1636, 1638, 1642, 1664, 1666, 1674, 1682, 1701, 1711, 1717, 1725, 1763, 1772, 1796, 1822, 1840, 1841, 1855, 1860, 1867, 1872, 1883, 1909, 1922, 1933, 1942, 1948, 1949, 1963, 1965, 1968, 1998, 2008, 2009, 2013]

Generating Sagittal View¶

based on this comment by @harshitsheoran which gives a decently usable saggital view¶

In [45]:
print(train_df.loc[41,:])
dcm_images = get_dcm_images(train_df['segment_path'][41]) # 0 fractures
dcm_np = np.array(dcm_images)
if train_df['reverse_required'][41] == True:
    dcm_np = dcm_np[::-1]
saggital_view = dcm_np[:,:,256]
plt.imshow(saggital_view, cmap='bone')

StudyInstanceUID                                                                            1.2.826.0.1.3680043.25704
patient_overall                                                                                                     0
C1                                                                                                                  0
C2                                                                                                                  0
C3                                                                                                                  0
C4                                                                                                                  0
C5                                                                                                                  0
C6                                                                                                                  0
C7                                                                                                                  0
total_fractures                                                                                                     0
segment_path              ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.25704
num_slices                                                                                                        257
nii_segments_path    ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.25704.nii
reverse_required                                                                                                False
Name: 41, dtype: object
Out[45]:
<matplotlib.image.AxesImage at 0x7f3d2c0345d0>
In [46]:
print(train_df.loc[60,:])
dcm_images = get_dcm_images(train_df['segment_path'][60]) # 4 fractures
dcm_np = np.array(dcm_images)
if train_df['reverse_required'][60] == True:
    dcm_np = dcm_np[::-1]
saggital_view = dcm_np[:,:,256]
plt.imshow(saggital_view, cmap='bone')

StudyInstanceUID                                                                            1.2.826.0.1.3680043.31077
patient_overall                                                                                                     1
C1                                                                                                                  0
C2                                                                                                                  0
C3                                                                                                                  1
C4                                                                                                                  1
C5                                                                                                                  1
C6                                                                                                                  1
C7                                                                                                                  0
total_fractures                                                                                                     4
segment_path              ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.31077
num_slices                                                                                                        672
nii_segments_path    ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.31077.nii
reverse_required                                                                                                False
Name: 60, dtype: object
Out[46]:
<matplotlib.image.AxesImage at 0x7f3d22c880d0>
In [47]:
print(train_df.loc[4,:])
dcm_images = get_dcm_images(train_df['segment_path'][4]) # 4 fractures
dcm_np = np.array(dcm_images)
if train_df['reverse_required'][4] == True:
    dcm_np = dcm_np[::-1]
saggital_view = dcm_np[:,:,256]
plt.imshow(saggital_view, cmap='bone')

StudyInstanceUID                                                                            1.2.826.0.1.3680043.1363
patient_overall                                                                                                    1
C1                                                                                                                 0
C2                                                                                                                 0
C3                                                                                                                 0
C4                                                                                                                 0
C5                                                                                                                 1
C6                                                                                                                 0
C7                                                                                                                 0
total_fractures                                                                                                    1
segment_path              ../input/rsna-2022-cervical-spine-fracture-detection/train_images/1.2.826.0.1.3680043.1363
num_slices                                                                                                       199
nii_segments_path    ../input/rsna-2022-cervical-spine-fracture-detection/segmentations/1.2.826.0.1.3680043.1363.nii
reverse_required                                                                                                True
Name: 4, dtype: object
Out[47]:
<matplotlib.image.AxesImage at 0x7f3d22d298d0>

The below code can be run to get all sagittal plane views¶

In [ ]:
# !mkdir saggital_views

# saggital_path = Path('./saggital_views')
# for uid, path, rev in tqdm(train_df.loc[:,['StudyInstanceUID','segment_path','reverse_required']].values):
#     dcm_images = get_dcm_images(path)
#     dcm_np = np.array(dcm_images)
#     if rev == True:
#         dcm_np = dcm_np[::-1]
#     saggital_view = dcm_np[:,:,256]
#     plt.imsave(saggital_path/(uid+'.png'),saggital_view,cmap='bone')

Save the training dataset with added columns and the dataset containing UIDs with segments¶

In [48]:
train_df.to_csv('train_new.csv')
uids_with_segments.to_csv('train_with_segments.csv')

Let us review a training sample data index animation¶

In [49]:
sample_index = train_df.iloc[99,:]
dcm_images = get_dcm_images(sample_index['segment_path'])
if sample_index['reverse_required'] == True:
    dcm_images.sort(reverse=True)
segments = get_nii_segments(sample_index['nii_segments_path'])
In [50]:
from matplotlib import animation, rc
rc('animation', html='jshtml')

fig, [ax1,ax2] = plt.subplots(1,2)
ax1.axis('off')
ax2.axis('off')
images = []
for i in tqdm(range(len(dcm_images))):
    im1 = ax1.imshow(dcm_images[i], animated=True, cmap='bone')
    im2 = ax2.imshow(segments[i,:,:], animated=True, cmap='bone')
    if i==0:
        ax1.imshow(dcm_images[i], cmap='bone')
        ax2.imshow(segments[i,:,:], cmap='bone')
    images.append([im1,im2])
        

ani = animation.ArtistAnimation(fig, images, interval=50, blit=True,
                                repeat_delay=1000)
plt.close()
ani

  0%|          | 0/271 [00:00<?, ?it/s]
Out[50]:

References¶

References thus far:¶

  • Discussions¶

    • Explaining Data and Submission in detail by @harshitsheoran¶

    this was very helpful. @harshitsheoran explained everything perfectly which gave me the confidence to try out this dataset

    • Order of Slices by @solverworld¶

    • Sorting the DICOM files by @abebe9849¶

    • Generating Saggital Views by @harshitsheoran¶

  • Notebooks¶

    • rsna-fracture-detection-dicom-images-explore by @andradaolteanu¶

    • rsna-fracture-detection-in-depth-eda by @samuelcortinhas¶

    • spine-fracture-eda-loading-dicom-3d-browse by @jirkaborovec¶

    • a-segmentation-is-in-reverse-order by @itsuki9180¶

Will continue to add more as referenced from different sources¶